Disclosure and Choice1
Elchanan Ben-Porath
2
Eddie Dekel
3
Barton L. Lipman4
First Draft
December 2014
1 We
thank Rick Green, Mark Machina, Phil Reny, and various seminar audiences for helpful
comments and the National Science Foundation, grant SES–0820333 (Dekel), and the US–Israel
Binational Science Foundation for support for this research.
2 Department of Economics and Center for Rationality, Hebrew University. Email: benporat@math.huji.ac.il.
3 Economics Department, Northwestern University, and School of Economics, Tel Aviv University. Email: dekel@northwestern.edu.
4 Department of Economics, Boston University. Email: blipman@bu.edu.
Abstract
An agent chooses among projects with random outcomes. His payoff is increasing in
the outcome and in an observer’s expectation of the outcome. With some probability, the
agent can disclose the true outcome to the observer. We show that choice is inefficient:
the agent favors riskier projects even with lower expected returns. If information can be
disclosed by a challenger who prefers lower beliefs of the observer, the chosen project is
excessively risky when the agent has better access to information, excessively risk–averse
when the challenger has better access, and efficient otherwise. We also characterize the
agent’s worst–case equilibrium payoff.
1
Introduction
Consider an agent who makes productive decisions and also decisions about how much to
disclose about the outcomes of these choices. The productive decisions are not observed
directly and the outcome of the choice is only observed after some delay. The agent’s
payoff depends on the outcome of the productive decisions but also on the beliefs of an
observer regarding the outcome prior to its observation. We give several examples of this
situation below.
Intuitively, the agent has an incentive to engage in excessive risk–taking. After all,
he can (at least to some extent for some period of time) hide bad outcomes and disclose
only good ones. This creates an option value which encourages risk–taking. We show
that this incentive harms the agent in the sense that he would be better off if he had no
control over information. The reason is that the agent always has an incentive to try to
choose a project that looks better than it is. In equilibrium, though, the observer cannot
be fooled, so the agent simply hurts himself. In particular, he would be better off if
he could commit to never disclosing anything or to any other “nonstrategic”1 disclosure
policy. We refer to the outcome that gives the best possible payoff to the agent as the
first best and show that this is the outcome when the agent cannot affect disclosure. We
also show that the agent’s utility loss relative to the first best can be “large” in a sense
to be made precise.
We now give examples of this setting.
First, consider the manager of a firm. His actions determine a probability distribution
over the firm’s profits. In the short run, he can choose to release privately observed
information about profits. The observer is the stock market whose beliefs about the
firm’s profits determine the stock price of the firm. The manager’s payoff is a convex
combination of the short–run and long–run stock price, where the latter is the realized
profits — the true value of the firm. Note that the manager’s utility function can be
identical to that of the stockholders in the firm, so the inefficiency we identify is not
due to a standard moral hazard problem. Here the first–best project is the one which
maximizes the expected value of the firm.
Second, suppose the agent is an incumbent politician and the observer is a representative voter. The productive activity chosen by the incumbent is a policy which affects
the utility of the voter. Before the outcome of the policy is observed, the incumbent
comes up for reelection. As part of his campaign, he may release information regarding
the progress of his policies. The probability the voter retains the incumbent is strictly
1
By “nonstrategic,” we mean any policy where the probability that information is disclosed is independent of the information being disclosed.
1
increasing in the voter’s beliefs about the utility he will receive from the incumbent’s
policy choice. One can think of this as retrospective voting or can assume that if the
incumbent is not reelected, his policy will be replaced by that of a challenger. The incumbent desires to be reelected and also cares about the true utility of the voters. In this
setting, the first–best project is that which maximizes the expected utility of the voters.
Third, an entrepreneur chooses a project which he will need to sell part of to a venture
capitalist at the interim stage. The funding he receives is increasing in the beliefs of
potential buyers about the value of the project. He may have private information he
could disclose at the interim stage regarding how well the project is progressing. Again,
the first–best project is the one with the highest expected value.
Fourth, consider a firm with multiple divisions, each of which could potentially head
up a prestigious project. The agent is the first division to have an opportunity to lead
and the observer is senior management. The agent has to decide among several ways to
try to achieve success on the project, where each method corresponds to a probability
distribution over profits from the project. The agent may have private information about
the progress of the project that he could disclose at the interim stage. If senior management believes the project has not been handled sufficiently well at the interim stage, it
transfers control to another division.
In some of these settings, it is natural to consider a challenger to the agent who might
also have access to information he may disclose. For example, in the case of an incumbent
politician, it is natural to suppose that a challenger running against him might be able
to disclose information about the incumbent’s policies. Similarly, in the example of a
firm deciding whether to retain the current project manager or opt for an alternative,
the alternative manager might have information about what is happening which he could
disclose.
We will show that in the extreme case where all information comes from the challenger,
the agent has an incentive to behave in a risk–averse manner. In effect, the option value
lies entirely with his opponent, so he wishes to minimize risk to minimize the value of
this (negative) option. When both the agent and the challenger can disclose, the effect
of disclosure on action choice depends on which is more likely to obtain information. If
the agent has more access to information in this sense than the challenger, excessively
risky decisions are made, while if the challenger has more access, then excessively risk–
averse choices result. Only when information is exactly balanced are production decisions
first–best.2
In all cases, we also characterize the worst possible equilibrium payoff for the agent
relative to the first–best payoff. For example, we show that the agent’s payoff can be as
2
If information is “close” to balanced, then production decisions are “close” to the first best.
2
low as 50% of the first–best payoff but cannot be any lower.
In the next section, we illustrate the basic ideas with a simple example. In Section
3, we give an overview of the most general version of the model. As we show in Section
6, the analysis of the general version can be reduced to the special cases where only the
agent has access to information to disclose and where only the challenger has such access.
In light of this and the fact that these special cases are simpler, we begin with them, in
Sections 4 and 5 respectively. Section 7 concludes.
The remainder of this introduction is a brief survey of the related literature. There
is a large literature on disclosure, beginning with Grossman (1981) and Milgrom (1981).
These papers established a key result which is useful for some of what follows. They consider a model where an agent wishes to persuade an observer, but only through disclosure
— the agent does not affect the underlying distribution over outcomes. They assume the
agent is known to have information and show that unraveling leads to the conclusion that
the unique equilibrium is for the agent to always disclose his information. Roughly, the
reasoning is that the agent with the best possible information will disclose, rather than
pool with any lower types. Hence the agent with the second–best possible information
cannot pool with the better information and so will also disclose, etc. Subsequent important contributions including Verrecchia (1983), Dye (1985a), Jung and Kwon (1988),
Fishman and Hagerty (1990), Okuno–Fujiwara, Postlewaite, and Suzumura (1990), Shin
(1994, 2003), Lipman and Seppi (1995), Glazer and Rubinstein (2004, 2006), Forges and
Koessler (2005, 2008), Archarya, DeMarzo, and Kremer (2011), and Guttman, Kremer,
and Skrzypacz (2013) add features to the model which block this unraveling result and
explore the implications. To explore the effect of disclosure on productive activities by
the agent, we also need a model of disclosure in which unraveling does not occur. We
use the approach initially developed by Dye (1985a) and Jung and Kwon (1988) for this
purpose.
While the literature on disclosure is large, relatively little attention has been paid to
the interaction of disclosure and production decisions and the papers that do consider
this take very different approaches from ours. Some papers consider “real effects” of
disclosure through its effect on the discloser’s competitors (Verrecchia (1983) or Dye
(1985b)) or effects that work through how disclosure affects the informativeness of stock
prices (Diamond and Verrecchia (1991), Bond and Goldstein (2014), or Gao and Liang
(2013)). While these do generate costs which can have effects on the firm’s productive
investments, they are very different sources of costs than the incentive effects we focus
on.
A few other papers consider incentive effects. A number of these follow Stein (1989)
in assuming that the manager may have an incentive to divert future cash flows to the
present in order to mislead the market about the long–run value of the firm. In this
3
setting, the nature of mandatory disclosure rules (e.g., the frequency of disclosure and
the kind of information which must be disclosed) have welfare implications through the
effect on the manager’s diversion of cash flows or other investment distortions. See,
for example, Kanodia and Mukherji (1996), Kanodia, Sapra, and Venugopalan (2004),
Edmans, Heinle, and Huang (2013), or Gigler, Kanodia, Sapra, and Venugopalan (2013).
The short–termism effect explored in these papers is similar to the inefficiency we consider
in that both approaches consider how the manager of a firm can manipulate what the
market observes. There are two key differences between the approaches. First, the
inefficiencies we identify are driven by strategic disclosure, a consideration unrelated to
the short–termism results. Second, our results are on inefficiencies in the riskiness of
investment while the short–termism literature concerns inefficiencies in its timing.3
A different approach to incentive effects is taken by Beyer and Guttmann (2012) who
consider a model in which disclosure interacts with investment and financing decisions.
Their paper is primarily focused on the signaling effects stemming from private information about the exogenous quality of investment opportunities. Thus both the nature and
source of the inefficiency are very different from what we consider.
2
Illustrative Example
We begin with an illustrative example to highlight the intuition of our results. This
example is for a special case of the environment, where the agent has no challenger and
cares only about the observer’s beliefs. We explain the model in more detail in the next
section, stating here only what is needed for the example. Specifically, we analyze the
perfect Bayesian equilibria of a three–stage game. In the first stage, the agent chooses
a project to undertake where a project corresponds to a lottery over outcomes in R+ .
In the second stage, with probability q1 , the agent receives evidence revealing the exact
realization from the project. If he receives evidence, he can either disclose it or withhold
it. (If he has no evidence, he cannot show anything.)
The observer does not see the project chosen by the agent or whether he has evidence;
the observer sees only the evidence, if any, which is presented. In the third stage, the
observer forms a belief b about the outcome of the project which equals the expectation
of the outcome conditional on all public information. Thus if evidence was presented
in the second stage, the observer’s belief must equal the outcome shown. The agent’s
3
These papers can be seen as part of a broader literature on moral hazard in corporate finance and
accounting. As in our paper, the manager, even if he represents the interests of current shareholders,
has an incentive to take actions to try to “fool” the market or other investors but, of course, is correctly
interpreted in equilibrium. As a result, he is worse off than if he could have committed to efficient choices
in the first place. See, for example, the risk shifting problem discussed in Jensen and Meckling (1976).
4
payoffs equal the observer’s belief, b.
Consider the following example. Assume q1 ∈ (0, 1), so the agent may or may not
have information. Also, assume that there are only two projects, F and G, where G is a
degenerate distribution yielding x = 4 with probability 1 and F gives 0 with probability
1/2 and 6 with probability 1/2. Recall that the agent’s ex ante payoff is the expectation
of the observer’s belief b. In equilibrium, the observer will make correct inferences about
the outcome of the project given what is or is not disclosed, so the expectation of the
observer’s belief must equal the expectation of x under the project chosen by the agent.
Hence if we have an equilibrium in which F is chosen, then the agent’s ex ante payoff
must be 3, while if we have an equilibrium in which G is chosen, the agent’s ex ante
payoff must be 4. In this sense, G is the best project for the agent, the one he would
commit himself to if he could. For this reason, we say G is the first–best project and that
4 is the agent’s first–best payoff.
Despite the fact that the agent would like to commit to G, it is not an equilibrium
for him to choose it. To see this, suppose the observer expects the agent to choose this
project. Then if the agent discloses nothing, the observer believes this is only because
the agent did not receive any information (an event with positive probability in the
hypothetical equilibrium as q1 < 1) and so believes x = 4. Given this, suppose the agent
deviates to project F . Since the project choice is not seen by the observer, the observer’s
beliefs cannot change in response. If the outcome of project F is observed by the agent to
be 0, he can simply not disclose this and the observer will continue to believe that x = 4. If
the outcome is observed to be 6, the agent can disclose this, changing the observer’s belief
to x = 6. Hence the agent’s payoff to deviating is (1−q1 )(4)+q1 [(1/2)(4)+(1/2)(6)] > 4.
So it is not an equilibrium for the agent to choose project G. One can show that if
0 < q1 ≤ 1/2, then the unique equilibrium in this example is for the agent to choose
project F .4 Thus the agent is worse off than in the first–best. His inability to commit
leads him to deviate from projects that are efficient but not “showy” enough. Since such
deviations are anticipated in equilibrium, he ends up choosing an inefficient project and
suffering the consequences.
In this example, the agent’s expected payoff as a proportion of his first–best payoff
is 3/4. An implication of Theorem 3 is that, for all q1 and all sets of feasible projects,
the agent’s equilibrium payoff must be at least half the first–best utility and that this
bound can be essentially achieved (that is, we can find parameters for which there is an
equilibrium payoff as close as we want to this bound).
4
If q1 is larger, the unique equilibrium is mixed.
5
3
Model
In this section, we present the most general version of the model and explain the basic
structure of equilibria. In the following sections, we discuss the inefficiencies of the
equilibrium.
Now the game has three players — the agent, the challenger, and the observer. As
in the example, there are three stages. In the first stage, the agent chooses a project
to undertake. Each project corresponds to a lottery over outcomes. The set of feasible
lotteries is denoted F where each F ∈ F is a (cumulative) distribution function over R+ .
For simplicity, we assume the supports of the feasible distributions are bounded from
below by 0 and from above by x̄. That is, we assume that there exists x̄ < ∞ such that
F (x̄) = 1 for all F ∈ F. We assume the set F is finite with at least two elements.5
In the second stage, there is a random determination of whether the agent or challenger has evidence demonstrating the outcome of the project. Let q1 denote the probability that the agent has evidence and q2 the probability that the challenger has evidence.
We assume that the events that the agent has evidence and that the challenger has evidence are independent of one another and that both are independent of the project chosen
by the agent and its realization.6 If a player has evidence, then he can either present it,
demonstrating conclusively what the outcome of the project is, or he can withhold it. If
he has no evidence, he cannot show anything. The decisions by the agent and challenger
regarding whether to show their evidence (if they have any) are made simultaneously.7
Neither the agent nor the challenger sees whether the other has evidence. The observer
does not see the project chosen by the agent nor whether he or the challenger has evidence
— the observer sees only the evidence, if any, which is presented and by whom.
In the third stage, the observer forms a belief b about the outcome of the project which
equals the expectation of x conditional on all public information.8 Thus if evidence was
presented in the second stage, the observer’s belief must equal the outcome shown since
evidence is conclusive.
Finally, the outcome of the project is realized and observed. The payoffs are as
5
The assumption that F is finite is a simple way to ensure equilibrium existence. It is not difficult
to allow for unbounded supports as long as all relevant expectations exist.
6
As shown in Section 6, our results do not rely on the first of these independence assumptions. We
use it only for notational convenience.
7
As will be clear from the analysis, the results also hold if the players move sequentially.
8
For expositional simplicity, we do not explicitly model the payoffs of the observer as they are irrelevant for the equilibrium analysis. Among other formulations, one could assume that the observer
chooses an action b and has payoff −(x − b)2 . Obviously, the observer would then choose b equal to the
conditional expected value of x. The examples in the introduction suggest various other payoff functions
for the observer.
6
follows. Let x be the realization of the project and b the observer’s belief in the third
stage. The agent’s payoff is αx + (1 − α)b where α ∈ [0, 1]. The challenger’s payoff is −b.
Because the challenger cannot affect x, the results would be the same if we assumed the
challenger’s payoff is βx + (1 − β)(−b) for β ∈ [0, 1), for example.
Note that the game is completely specified by a feasible set of projects F and the
values of α, q1 , and q2 . For this reason, we sometimes write an instance of this game as
a tuple (F, α, q1 , q2 ). Throughout, we consider perfect Bayesian equilibria.9
In the remainder of this section, we do the following. First, we discuss the benchmark
case where the information seen by the observer cannot be affected by the agent or
challenger — where it is entirely exogenous. As we will show, this case generates the
first–best outcome, which is the outcome which maximizes the agent’s expected payoff
over all feasible projects. Second, we discuss the structure of equilibria in this game
more generally to set up our detailed discussion of the inefficiencies of equilibria in the
following sections.
3.1
Benchmark
First, we consider the benchmark case where the information seen by the observer is not
strategically determined. In other words, suppose the observer sees the realization of the
project at stage 2 with probability q ∈ [0, 1] and that the agent and challenger cannot
affect whether the observer sees this information.
Except for the degenerate case where α = q = 0, the optimal project choice by the
agent is any project F which maximizes EF (x) where EF denotes the expectation with
respect to the distribution F . We refer to such a project F as a first–best project.
To see why the agent chooses a first–best project, fix an equilibrium. Let x̂ denote the
belief of the observer if he does not see any evidence. Then if the agent chooses project
F , his expected payoff is
αEF (x) + (1 − α) [qEF (x) + (1 − q)x̂] .
Obviously, if α = q = 0, then the agent’s payoff is x̂, regardless of the F he chooses, so he
is indifferent over all projects. Otherwise, his payoff is maximized by maximizing EF (x).
To be more precise, choosing any such F strictly dominates choosing any project with
a strictly lower expectation. (The degeneracy of the case where α = q = 0 will appear
again below.)
9
Specifically, we use what Mas-Colell, Whinston, and Green (1995) refer to as weak perfect Bayesian
equilibrium. Our results would not change if we used a stronger notion such as Kreps and Wilson’s
(1982) sequential equilibrium.
7
As the example in Section 2 showed, equilibria are typically not first–best when
disclosure is chosen by the agent strategically. If the observer expects the agent to
choose a first–best project, he may have an incentive to deviate to a less efficient project
which has a better chance of a very good outcome, preventing his choice of the first–best
from being an equilibrium. Hence he ends up choosing a project with a lower expected
value and is worse off as a result.
3.2
Equilibrium
Now we turn to the general structure of equilibria in this model. So suppose we have an
equilibrium where the agent uses a mixed strategy σ where σ(F ) is the probability the
agent chooses project F . Again, let x̂ denote the belief of the observer if he is not shown
any evidence at stage 2. If q1 and q2 are both strictly less than 1, then this information
set must have a strictly positive probability of being reached.
Given x̂, it is easy to determine the optimal disclosure strategies for the agent and
the challenger. First, suppose the agent obtains proof that the outcome is x where
x > x̂. In this case, the agent will disclose the outcome in any equilibrium, regardless
of the strategy of the challenger. Clearly, if the probability the challenger would reveal
this information is less than 1, then the agent is strictly better off revealing than not
revealing. So suppose the challenger reveals this information with probability 1 — that
is, q2 = 1 and the challenger’s strategy given x is to disclose it. Since the challenger would
not want to reveal this information, the only way this could be optimal for the challenger
is if the agent is also disclosing it, rendering the challenger indifferent between disclosing
and not. Hence, either way, the agent must disclose this information with probability 1.
Similar reasoning shows that if the challenger obtains proof that the outcome is x where
x < x̂, then the challenger discloses this with probability 1.
So suppose the agent obtains proof that the outcome is x < x̂. Similar reasoning to
the above shows that he hides this information in equilibrium except in the trivial case
where q2 = 1. When q2 = 1, the challenger will necessarily also have this information.
From the above, we know the challenger will disclose it. Hence in this case, the agent
is indifferent between disclosing and not. In short, if the agent’s disclosure decision
matters, then he does not disclose in this situation. For simplicity, we simply focus on
the equilibrium where the agent never discloses when x < x̂. Similar reasoning shows
that we can also assume without loss of generality that the challenger never discloses
when x > x̂.
One can show that the equilibrium is entirely unaffected by the disclosure choices
when x = x̂, so for simplicity we assume both the agent and challenger disclose in this
8
situation.10
In light of this, we can write the agent’s payoff as a function of the project F and x̂
as
î
VA (F, x̂) =αEF (x) + (1 − α) (1 − q1 )(1 − q2 )x̂ + q1 (1 − q2 )EF max{x, x̂}
ó
(1)
+ q2 (1 − q1 )EF min{x, x̂} + q1 q2 EF (x) .
We can complete the characterization of equilibria as follows. First, given x̂, we have
VA (F, x̂) = max VA (G, x̂) for all F such that σ(F ) > 0.
G∈F
That is, the agent’s mixed strategy is optimal given the disclosure behavior described
above and the observer’s choice of x̂.
Second, given σ, x̂ must be the expectation of x conditional on no evidence being
presented and given the disclosure strategies and the observer’s belief that the project
was chosen according to distribution σ. The most convenient way to state this is to use
the law of iterated expectations to write it as
X
F ∈F
σ(F )EF (x) =
X
î
σ(F ) (1 − q1 )(1 − q2 )x̂ + q1 (1 − q2 )EF max{x, x̂}
F ∈F
(2)
ó
+ q2 (1 − q1 )EF min{x, x̂} + q1 q2 EF (x) .
The left–hand side is the expectation of x given the mixed strategy used by the agent in
selecting a project. The right–hand side is the expectation of the observer’s expectation
of x given the disclosure strategies and the agent’s mixed strategy for selecting a project.
Equation (2) implies that the agent’s equilibrium expected payoff is F ∈F σ(F )EF (x).
Thus the agent’s payoff in any equilibrium must be weakly below the first–best payoff.
P
Also, if α = q1 = q2 = 0, then VA (F, x̂) = x̂. In this case, the agent’s actions do
not affect his payoff, so he is indifferent over all projects. Henceforth, we refer to a
game (F, α, q1 , q2 ) with α = q1 = q2 = 0 as degenerate and call the game nondegenerate
otherwise.
10
It is obvious that a player’s choice when he observes x = x̂ is irrelevant if this is a measure zero
event. However, even with discrete distributions, this remains true. First, obviously, a player’s payoff
is unaffected by what he does when indifferent. Second, if either the agent or challenger is indifferent,
the other player is as well, so the agent’s choice doesn’t affect the challenger or conversely. Finally, the
indifferent player’s choice does not affect the observer’s posterior beliefs since this is a matter of whether
we include a term equal to the average in the average or not — it cannot affect the calculation.
9
4
Agent Only
In this section, we focus on the case where the challenger is effectively not present.
Specifically, we consider the model of the previous section for the special case where
q2 = 0. This is of interest in part because there is no obvious counterpart of the challenger
in some natural examples which otherwise fit the model well. Also, as we will see in
Section 6, the general model reduces either to this special case or the special case discussed
in the next section where only the challenger may have information.
When q2 = 0, equation (1) defining VA (F, x̂) reduces to
VA (F, x̂) = αEF (x) + (1 − α)[(1 − q1 )x̂ + q1 EF max{x, x̂}].
(3)
Thus the agent chooses the project F to maximize EF [αx + (1 − α)q1 max{x, x̂}] for a
certain value of x̂. If x̂ were exogenous and we simply considered αx+(1−α)q1 max{x, x̂}
to be the agent’s von Neumann–Morgenstern utility function, we would conclude that
the agent is risk loving since this is a convex function of x (as long as (1 − α)q1 > 0).
The results we show below build on this observation, making more precise the way this
incentive to take risks is manifested in the agent’s equilibrium choices.
To clarify the sense in which the agent’s choices are risk seeking, we first recall some
standard concepts.
Definition 1. Given two distributions F and G over R+ , G dominates F in the sense
of second–order stochastic domination, denoted G SOSD F , if for all z ≥ 0,
Z z
0
F (x) dx ≥
Z z
0
G(x) dx.
We say that F is riskier than G if G SOSD F and EF (x) = EG (x).
It is well–known that if G SOSD F , then every risk averse agent prefers G to F . If F
is riskier than G, then every risk–loving agent prefers F to G and every risk neutral agent
is indifferent between the two. The reason that the mean condition has to be added for
the second two comparisons is that if G SOSD F , then the mean of G must be weakly
larger than the mean of F . Clearly, if it is strictly larger, then G could be better than F
even for a risk–loving agent.
Our first result on risk taking uses a stronger notion of riskier.
Definition 2. Given two distributions F and G over [a, b], G strongly dominates F in the
sense of second–order stochastic domination, denoted G SSOSD F , if for all z ∈ (a, b),
Z z
0
F (x) dx >
Z z
0
G(x) dx.
We say that F is strongly riskier than G if G SSOSD F and EF (x) = EG (x).
10
One can show that if F is strongly riskier than G, then for every continuous and
increasing utility function u with uniformly bounded directional derivatives, F yields
strictly higher expected utility than G if u is convex and not linear, while G yields
strictly higher expected utility than F if u is concave and not linear.
Theorem 1. Suppose q2 = 0. Suppose there are distributions F, G ∈ F such that F
is strongly riskier than G. Then if α < 1 and q1 ∈ (0, 1), there is no pure strategy
equilibrium in which the agent chooses G.11
Proof. Suppose to the contrary that it is a pure equilibrium for the agent to choose G.
Then the payoff to G must exceed the payoff to F . Using equation (3), this implies
αEG (x) + (1 − α)q1 EG max{x, x̂} ≥ αEF (x) + (1 − α)q1 EF max{x, x̂}.
Since F is strongly riskier than G, they have the same mean so, given α < 1 and q1 > 0,
this reduces to
EG max {x, x̂} ≥ EF max {x, x̂} .
Note that
EF max{x, x̂} = F (x̂)x̂ +
Z x̄
x dF (x).
x̂
Integration by parts shows that
Z x̂
0
F (x) dx = F (x̂)x̂ −
Z x̂
0
x dF (x) = F (x̂)x̂ − EF (x) +
so
EF max{x, x̂} = EF (x) +
Z x̂
0
Z x̄
x dF (x),
x̂
F (x) dx.
Hence we must have
EG (x) +
Z x̂
0
G(x) dx ≥ EF (x) +
Z x̂
0
F (x) dx.
Again, since F is strongly riskier than G, we have EG (x) = EF (x) implying
Z x̂
0
G(x) dx ≥
Z x̂
0
F (x) dx.
Since F is strongly riskier than G, this implies that x̂ ∈
/ (a, b). Hence, in particular,
either x̂ is strictly outside the support of G or is either the upper or lower bound of the
support.
11
It is worth noting that this result also holds in a model of project choice with disclosure modeled as
in Verrecchia (1983) if the cost of disclosure is small enough.
11
From equation (2), x̂ must satisfy
EG (x) = (1 − q1 )x̂ + q1 EG max{x, x̂}.
If x̂ is less than or equal to the lower bound of the support of G, then this equation
says EG (x) = (1 − q1 )x̂ + q1 EG (x). Hence, given the assumption that q1 < 1, we obtain
x̂ = EG (x), a contradiction unless G is degenerate at x̂. If x̂ is greater than or equal
to the upper bound of the support of G, then this equation says EG (x) = x̂, again a
contradiction unless G is degenerate at x̂.
So suppose G is degenerate at x̂. Since F =
6 G, F cannot be degenerate at x̂. Since
EF (x) = EG (x), x̂ must be in the interior of the support of F . But then x̂ ∈ (a, b), a
contradiction.
Our next result uses weaker hypotheses — comparing distributions using riskiness
rather than strong riskiness, allowing mixed equilibria, and not imposing parameter
restrictions other than non–degeneracy on α and q1 . Consequently, the conclusion is
weaker as well. Specifically, we show that if there are two distributions in F which can
be compared in terms of riskiness, the agent never chooses the less risky of the two if
the difference is ever relevant. To understand this result, note that the agent’s objective
function is piecewise linear, not strictly convex. Hence there are certain comparisons of
lotteries where the difference in risk is irrelevant to the agent. To make this last part of
the statement precise requires another definition.
Definition 3. F is equilibrium–indifferent to G if for every equilibrium in which G
receives positive probability, there is another equilibrium in which the agent’s mixed strategy is unchanged except the probability he played G previously is now moved to F , the
observer’s strategy is unchanged, and the agent’s expected payoff is unchanged.
It is not hard to show that if F is equilibrium–indifferent to G, then λF + (1 − λ)G
is also equilibrium–indifferent to G for all λ ∈ (0, 1). In other words, if F is equilibrium–
indifferent to G, then the agent makes no distinction between F and G whatsoever and
the observer’s behavior makes no distinction.
Theorem 2. Suppose q2 = 0. Suppose there are distributions F, G ∈ F such that F is
riskier than G. Then for any equilibrium of any nondegenerate game, either the agent
puts zero probability on G or else F is equilibrium–indifferent to G.
Proof. Fix distributions F and G with F riskier than G. Fix an equilibrium in which G
is in the support of the agent’s mixed strategy and define x̂ to be the observer’s response
if no evidence is presented in the equilibrium. Since G is given positive probability, we
must have
αEG (x) + (1 − α)q1 EG max{x, x̂} ≥ αEF (x) + (1 − α)q1 EF max{x, x̂}.
12
Since the game is nondegenerate, either α > 0 or q1 > 0 or both. Hence if (1 − α)q1 = 0,
we must have α > 0 so the agent maximizes EF (x) and the result follows. So for
the remainder, assume (1 − α)q1 > 0. The same integration by parts and the same
substitutions as used in the proof of Theorem 1 imply
Z x̂
0
G(x) dx ≥
Z x̂
0
F (x) dx.
F riskier than G implies the reverse weak inequality, so
Z x̂
0
G(x) dx =
Z x̂
0
F (x) dx,
implying
EG max{x, x̂} = EF max{x, x̂}.
Change the agent’s strategy by switching the probability he plays G to playing F . If x̂
does not change, his new strategy is still a best reply. From equation (2), the appropriate
x̂ can be defined by
(1 − q1 )x̂ + q1
X
σ(F 0 )EF 0 max{x, x̂} =
F 0 ∈F
X
σ(F 0 )EF 0 (x).
F 0 ∈F
Since EF (x) = EG (x) and EF max{x, x̂} = EG max{x, x̂}, we see that x̂ does not change.
Hence this is an equilibrium. Clearly, the agent obtains the same expected payoff. So F
is equilibrium–indifferent to G.
Theorems 1 and 2 compare distributions with the same means, but it is easy to see
that, in general, the agent will accept a lower mean in order to obtain more risk.
As an extreme illustration, we generalize the example of Section 2 as follows. Suppose
α = 0 and let G be a degenerate distribution yielding x∗ with probability 1. There is a
pure strategy equilibrium in which the agent chooses G if and only if there is no other
feasible distribution that has any chance of producing a larger outcome. That is, this
is an equilibrium iff there is no F ∈ F with F (x∗ ) < 1. The conclusion that G is an
equilibrium if F (x∗ ) = 1 for all F ∈ F is obvious, so consider the converse. Suppose we
have an equilibrium in which the agent chooses G but F (x∗ ) < 1. Because the agent is
expected to choose G, we have x̂ = x∗ . But then the agent could deviate to F and with
some (perhaps very small probability) will be able to show a better outcome than x∗ ,
yielding a payoff strictly above x∗ . If he cannot, he shows nothing and receives payoff
x∗ . Hence his expected payoff must be strictly larger than x∗ , a contradiction. Note that
the mean of x under F could be arbitrarily smaller than the mean under G.
While the mean of the distribution to which the agent deviates can be arbitrarily
smaller than the mean of G, this does not say that the agent’s payoff loss in equilibrium is
13
arbitrarily large. Below, we give tight lower bounds on the ratio of the agent’s equilibrium
payoff to his best feasible payoff. One simple implication of this result is that, except in
the degenerate case where α = q1 = 0, the agent’s equilibrium payoff must always be at
least half of his first–best payoff.
The more general result characterizes the ratio of the worst equilibrium payoff for the
agent to the first–best payoff.12 More precisely, given a game (F, α, q1 , q2 ), let
U F B (F) = max EF (x).
F ∈F
So U F B is the first–best payoff for the agent. Let U(F, α, q1 , q2 ) denote the set of equilibrium payoffs for the agent in the game. We will give a function R(α, q1 , q2 ) with the
following properties. First, for every F, for every U ∈ U(F, α, q1 , q2 ),
U ≥ R(α, q1 , q2 )U F B (F).
That is, R(α, q1 , q2 ) is a lower bound on the proportion of the first–best payoff that can
be obtained in equilibrium — i.e., on U/U F B for any equilibrium for any feasible set F.
Second, this bound is tight in the sense that for every ε > 0, there exists F and
U ∈ U(F, α, q1 , q2 ) such that
U ≤ R(α, q1 , q2 )U F B (F) + ε.
We therefore sometimes refer to R as the “worst–case payoff” for the agent.
In this section, we focus on games with q2 = 0, so we only characterize the function
for this special case here, giving the more general characterization later.
Specifically, we have the following result.13
Theorem 3. For any nondegenerate game, we have
R(α, q1 , 0) =
α + (1 − α)q1
.
α + (1 − α)q1 (2 − q1 )
Also, R(0, 0, 0) = 0. Hence for α > 0,
√
1+ α
min R(α, q1 , 0) =
.
q1 ∈[0,1]
2
12
This is essentially the inverse of what is sometimes called the Price of Anarchy. See, for example,
Koutsoupias and Papadimitriou (1999), who coined the term, or Roughgarden (2005).
13
The exact statements of the lower bounds in Theorems 3 and 5 exploit our normalization that
the outcome from any project is non–negative. However, it is straightforward to adapt these bounds
to the more general case where there is some (not necessarily positive) lower bound for all supports.
Specifically, suppose x is a lower bound for all supports. When x = 0, our theorems characterize a
function R such that U ≥ RU F B and this bound is tight. When x 6= 0, what we are establishing is that
U ≥ RU F B + (1 − R)x and that this bound is tight. We thank Bruno Strulovici for raising this issue.
14
We offer several comments on this result. First, there is a discontinuity in the function
R at the degenerate case where α = q1 = q2 = 0. To see this, note from the characterization of the minimum over q1 that R(α, q1 , 0) ≥ 1/2 if α > 0, but R(0, 0, 0) = 0.
To understand this discontinuity, note that when q1 = q2 = 0, there is no information
that will be revealed to the observer at stage 2. When α = 0, the only thing the agent
cares about is the observer’s belief. Since no information will be revealed to the observer,
the agent cannot do anything to affect the only thing he cares about. In particular, for
any F ∈ F, it is an equilibrium for the agent to choose F since no deviation from this
F will change his expected payoff. Consequently, our remaining remarks focus on the
nondegenerate case.
Second, it is easy to see that R(α, q1 , 0) is increasing in α and equals 1 at α = 1.
Hence, as one would expect, if α = 1, we obtain the first–best. In this case, the agent
does not care about the observer’s belief, only the true realization of x, and so is led to
maximize it (in expectation).
Third, it is not hard to show that R(α, q1 , 0) is not monotonic in q1 except when
α = 0 or (trivially)
1. Specifically, given any α, the value of q1 which minimizes the
√ α =√
bound is q1 = α/[1 + α], which is interior for any α ∈ (0, 1).
This non–monotonicity stems from the fact that when α > 0, we obtain the first–best
at both q1 = 0 and at q1 = 1. That is, R(α, 0, 0) = R(α, 1, 0) = 1 for all α > 0. When
q1 = 0, the agent cannot influence the observer’s beliefs and so cares only about the
true value of x. Hence he chooses the project which maximizes its expectation. When
q1 = 1, he is known to always have information. So the standard unraveling argument
implies that he must reveal the information always. Hence he cannot be strategic about
disclosure and therefore will again maximize the expected value of x.
Figure 1 illustrates Theorem 3. It shows R(α, q1 , 0) as a function of q1 for various
values of α.
The proof of Theorem 3 is a little tedious and so is relegated to the Appendix. To
provide some intuition, we prove a simpler result here, namely, that for α = 0, the agent’s
payoff must be at least half the first–best in any nondegenerate game. That is, we prove
the last statement of the theorem for α = 0.
So fix any feasible set of projects F and any q1 ∈ (0, 1]. Fix any equilibrium
strategy σ for the agent and any project F in the support of σ which has the
expected value of x across projects in the support. Fix the x̂ of the equilibrium.
be any first–best project. As seen in the proof of Theorems 1 and 2, q1 > 0, F
support of σ, and the optimality of σ imply
EF max{x, x̂} ≥ EG max{x, x̂}
15
mixed
lowest
Let G
in the
R(α, q1 , 0)
1.0
α = 0.7
α = 0.4
0.75
α = 0.15
α=0
0.5
0.25
0.25
0.5
0.75
1.0
q1
Figure 1: “Worst Case” as a Function of q1 .
or
EF (x) +
Z x̂
0
F (x) dx ≥ EG (x) +
Z x̂
0
G(x) dx.
Since F (x) ≤ 1 and G(x) ≥ 0, this requires
EF (x) + x̂ ≥ EG (x).
(4)
From equation (2),
X
σ(F 0 )EF 0 (x) = (1 − q1 )x̂ + q1
F 0 ∈F
X
σ(F 0 )EF 0 max{x, x̂}.
F 0 ∈F
Since EF 0 max{x, x̂} ≥ EF 0 (x), we see that
X
σ(F 0 )EF 0 (x) ≥ x̂.
F 0 ∈F
Also, by our assumption that F is one of the projects with the lowest mean in the support,
we have
X
σ(F 0 )EF 0 (x) ≥ EF (x).
F 0 ∈F
Hence equation (4) implies
2
X
σ(F 0 )EF 0 (x) ≥ EF (x) + x̂ ≥ EG (x).
F 0 ∈F
16
So the agent’s payoff
claimed.
P
F 0 ∈F
σ(F 0 )EF 0 (x) must be at least half of the first–best payoff, as
To show that this bound is approximately achievable, consider the following example.
Let α = 0. Suppose F = {F, G} where F is a discrete distribution putting probability
1 − p on 0 and p on 1/p for some p ∈ (0, 1), so EF (x) = 1. Let G be a degenerate
distribution giving probability 1 to x = x∗ . We construct an equilibrium where F is
chosen by the agent, so the agent’s equilibrium payoff, U , is 1. We focus on the case
where x∗ > 1, so U F B = x∗ . If the observer expects the agent to choose F with probability
1, then by equation (2), x̂ solves
(1 − q1 )x̂ + q1 [(1 − p)x̂ + 1] = 1
so
x̂ =
1 − q1
.
1 − q1 p
This is an equilibrium iff EG max{x, x̂} ≤ EF max{x, x̂} or
max{x∗ , x̂} ≤ (1 − p)x̂ + 1
(1 − p)(1 − q1 )
2 − q1 − p
=
+1=
.
1 − q1 p
1 − q1 p
It is easy to show that x̂ < 1 while, by assumption, x∗ > 1. So we have an equilibrium iff
x∗ ≤
2 − q1 − p
.
1 − q1 p
Let x∗ equal the right–hand side. Then we have an equilibrium where the agent’s payoff
is 1, but the first–best payoff is x∗ . By taking q1 and p arbitrarily close to 0, we can make
x∗ arbitrarily close to 2, so the agent’s payoff is arbitrarily close to half the first–best
payoff.
The implication of Theorem 3 that the worst–case payoffs are increasing as the agent
cares more about the true x and less about the observer’s belief b is intuitive, but it is
important to note that this result does not carry over to equilibrium payoffs in general.
In Appendix C, we give an example which illustrates several senses in which equilibrium
payoffs can decrease as α increases for fixed F. In the example, there is a mixed strategy
equilibrium with payoffs that are decreasing in α. Also, this equilibrium is the worst
equilibrium for the agent for some parameters, showing that the worst equilibrium payoff
for a fixed F can decrease with α. Finally, the payoff in the worst pure strategy equilibrium is also decreasing in α for a certain range, showing that this result is not an artifact
related to mixed equilibria.
17
5
Challenger Only
In this section, we consider the case where q1 = 0 and q2 may be strictly positive. In this
case, the agent’s payoff as a function of x̂ and his chosen project F is
VA (F, x̂) = αEF (x) + (1 − α) [(1 − q2 )x̂ + q2 EF min{x, x̂}] .
Analogously to our discussion in Section 4, we see that given x̂, it is as if the agent has a
von Neumann–Morgenstern utility function of αx + (1 − α)q2 min{x, x̂}. If (1 − α)q2 > 0,
this function is concave, so the agent’s choices are effectively risk averse. This gives the
following analog to Theorem 2.14
Theorem 4. Suppose q1 = 0. Suppose there are distributions F, G ∈ F such that F is
riskier than G. Then for any equilibrium of any nondegenerate game, either the agent
puts zero probability on F or else G is equilibrium–indifferent to F .
Proof. Fix F and G as above and suppose we have an equilibrium where F is in the
support of the agent’s mixed strategy. Clearly, we must have
αEF (x) + (1 − α)q2 EF min{x, x̂} ≥ αEG (x) + (1 − α)q2 EG min{x, x̂}.
Since the game is nondegenerate, either α > 0 or q2 > 0 or both. If (1−α)q2 = 0, the agent
chooses the project to maximize EF (x) and the result holds. So assume (1 − α)q2 > 0.
Since F is riskier than G, we have EF (x) = EG (x). Hence,
EF min{x, x̂} ≥ EG min{x, x̂}.
Note that
min{a, b} + max{a, b} = a + b,
so
EF min{x, x̂} = EF (x) + x̂ − EF max{x, x̂}.
Hence we must have
EF (x) − EF max{x, x̂} ≥ EG (x) − EG max{x, x̂}.
We showed earlier that EF max{x, x̂} = EF (x) +
Z x̂
0
14
G(x) dx ≥
Z x̂
0
R x̂
0
F (x) dx. Substituting,
F (x) dx.
It is straightforward to give an analog for Theorem 1 as well.
18
Since F is riskier than G, this must hold with equality. Change the agent’s strategy by
switching the probability he plays F to playing G. If x̂ does not change, his new strategy
is still a best reply. From equation (2), x̂ can be defined by
(1 − q2 )x̂ + q2
X
σ(F 0 )EF 0 min{x, x̂} =
F 0 ∈F
X
σ(F 0 )EF 0 (x).
F 0 ∈supp(σ)
Since EF (x) = EG (x) and EF min{x, x̂} = EG min{x, x̂}, we see that x̂ does not change.
Hence this is an equilibrium. Clearly, the agent obtains the same expected payoff.
We can also characterize R for this case. More specifically, we have the following
analog to Theorem 3:
Theorem 5. For all nondegenerate games, we have
R(α, 0, q2 ) =
α
.
α + (1 − α)q2
Hence for α > 0,
min R(α, 0, q2 ) = α
q2 ∈[0,1]
and for q2 > 0,
min R(α, 0, q2 ) = 0.
α∈[0,1]
Figure 2 illustrates this result. It shows R(α, 0, q2 ) as a function of q2 for the same
values of α as used in Figure 1.
Theorem 5 has some features in common with Theorem 3. In particular, both results
show that the outcome must be first–best when α = 1 or when α > 0 and there is
zero probability of disclosure (i.e., q2 = 0). In both cases, the worst case improves as α
increases.
On the other hand, this result also shows several differences from Theorem 3. First,
while there is a discontinuity, it is somewhat different from the discontinuity noted previously. To be specific, for q1 > 0,
R(0, q1 , 0) =
1
1
≥ ,
2 − q1
2
while R(0, 0, 0) = 0. Here we see that
R(0, 0, q2 ) = 0
for all q2 . Hence the function R(0, q1 , 0) is discontinuous in q1 at q1 = 0 but the function
R(0, 0, q2 ) is not discontinuous in q2 at q2 = 0. On the other hand, there is a discontinuity
at (0, 0, 0) since R(α, 0, 0) = 1 for all α > 0.
19
R(α, 0, q2 )
1.0
α = 0.7
0.75
α = 0.4
0.5
α = 0.15
0.25
α=0
0.25
0.5
0.75
1.0
q2
Figure 2: “Worst Case” as a Function of q2 .
Second, this result implies that the worst case over α when q1 > 0 and q2 = 0 is
better than the worst case when q1 = 0 and q2 > 0. In the former case, we have R = 1/2,
while in the latter, we have R = 0. Since the lower bound is zero and payoffs are non–
negative, this implies that in the case where only the challenger speaks, the agent could
be arbitrarily worse off than at the first–best.
Third, recall that for α ∈ (0, 1), the worst case payoff in Theorem 3 was first decreasing, then increasing in q1 , equalling the first–best at both q1 = 0 and q1 = 1. Here the
worst case is always decreasing in q2 . In particular, we obtain the first–best at q2 = 0
but not at q2 = 1. This may seem unintuitive since at q2 = 1, the challenger is known to
have information and therefore the standard unraveling argument would seem to suggest
he must reveal it. Hence, one expects, it is as if the observer always saw the true x and
so the outcome would seem to necessarily be first–best.
To understand why we do not necessarily obtain the first best at q2 = 1, consider
the following example. Suppose q2 = 1 and that F = {F, G} where F gives 1 with
probability 1/2 and 3 with probability 1/2, while G gives 0 with probability 1/2 and 100
with probability 1/2. Obviously, G is the first–best project. But there is an equilibrium
in which the agent chooses F if α is not too large. To see this, consider the case where
α = 0. Suppose F is the project the observer expects the agent to choose. Then if the
20
challenger presents no evidence, the observer believes the outcome to have been 3 since
this is the worst possible outcome for the challenger under F . Because of this, the agent
has no incentive to deviate to G. If he does deviates and the outcome is 0, the challenger
can show this and the agent is hurt. If the outcome is 100, the challenger can hide this
and the observer thinks the outcome was 3. Thus the agent’s expected payoff to the
deviation is (1/2)(0) + (1/2)(3) < 2, so the agent prefers F . For small enough α > 0, the
same conclusion will follow.
Intuitively, it is true that if the challenger always learns the outcome of the project,
we get unraveling and all information is revealed along the equilibrium path — i.e., when
the agent chooses the equilibrium project. We do not necessarily get unraveling if the
agent deviates to an unexpected project and this creates the possibility of inefficient
equilibria.
On the other hand, the efficient outcome is also an equilibrium.15
Theorem 6. For any α, if q1 = 0 and q2 = 1, then there is an equilibrium in which the
agent chooses the first–best project.
The proof of this is straightforward. Suppose the agent is expected to choose F where
EF (x) ≥ EG (x) for all G ∈ F. Let x∗ denote the supremum of the support of F and set
x̂ = x∗ . That is, assume that if the challenger does not reveal x, the observer believes
the realization is the largest possible value under F . It is easy to see that this is what
unraveling implies given that the agent chooses F . So this is an equilibrium as long as
the agent has no incentive to deviate to a different project. By choosing F , the agent’s
payoff is EF (x). If he deviates to any other feasible project G, his expected payoff is
αEG (x) + (1 − α)EG min{x, x∗ } ≤ EG (x) ≤ EF (x).
So the agent has no incentive to deviate.
6
Agent and Challenger
Now we consider the case where both the agent and the challenger may have information
to disclose in the second stage. The following result shows that the analysis reduces to
either the case where only the agent has evidence or the case where only the challenger
has evidence, depending on whether q1 or q2 is larger.
15
It is also worth noting that the efficient outcome is the only equilibrium if all projects have the same
support. We thank Georgy Egorov for pointing this out.
21
Theorem 7. Fix (F, α, q1 , q2 ). If q1 ≥ q2 , then the set of equilibria is the same as for
the game (F, α̂, q̂1 , 0) where
α̂ = α + (1 − α)q2
and
q1 − q 2
.
1 − q2
If q1 ≤ q2 , then the set of equilibria is the same as for the game (F, α̂, 0, q̂2 ) where
q̂1 =
α̂ = α + (1 − α)q1
and
q̂2 =
q2 − q 1
.
1 − q1
Corollary 1. For any nondegenerate game with q1 = q2 , the outcome is first–best.
To see why Theorem 7 implies the corollary, suppose we have a nondegenerate game,
so it is not the case that α = q1 = q2 = 0. By Theorem 7, if q1 = q2 , the outcome is the
same in the game with α̂ = α + (1 − α)q2 > 0 and q̂1 = q̂2 = 0. As shown in Theorem 3,
the outcome must be first–best in this case.
Proof of Theorem 7. Fix (F, α, q1 , q2 ) and an equilibrium. Let x̂ be the observer’s belief
if no evidence is presented. First, assume q1 ≥ q2 . Recall that the agent chooses F to
maximize
î
αEF (x) + (1 − α) (1 − q1 )(1 − q2 )x̂ + q2 (1 − q1 )EF min{x, x̂}
ó
+ q1 (1 − q2 )EF max{x, x̂} + q1 q2 EF (x) .
Note that
EF min{x, x̂} + EF max{x, x̂} = EF [min{x, x̂} + max{x, x̂}] = EF (x) + x̂.
Hence
EF min{x, x̂} = EF (x) + x̂ − EF max{x, x̂}.
(5)
Substituting, we can rewrite the agent’s payoff as
[α + (1 − α)q2 ]EF (x) + (1 − α) [(1 − q1 )x̂ + (q1 − q2 )EF max{x, x̂}] .
(6)
Let α̂ = α + (1 − α)q2 , so 1 − α̂ = (1 − α)(1 − q2 ). We can rewrite the above as
ñ
ô
1 − q1
q1 − q 2
α̂EF (x) + (1 − α̂)(1 − α)
x̂ +
EF max{x, x̂} .
(1 − α)(1 − q2 )
(1 − α)(1 − q2 )
Let q̂1 = (q1 − q2 )/(1 − q2 ), so 1 − q̂1 = (1 − q1 )/(1 − q2 ) Then this is
α̂EF (x) + (1 − α̂)[(1 − q̂1 )x̂ + q̂1 EF max{x, x̂}].
22
This is exactly the agent’s payoff when the observer’s inference in response to no evidence is x̂ in the game (F, α̂, q̂1 , 0). Hence the agent’s best response to x̂ in the game
(F, α, q1 , q2 ) is the same as in the game (F, α̂, q̂1 , 0).
To see that the observer’s belief given a mixed strategy by the agent also does not
change, note that we can rewrite equation (2) as
X
σ(F )EF (x) =
F ∈F
X
¶
î
σ(F ) αEF (x) + (1 − α) (1 − q1 )(1 − q2 )x̂
F ∈F
ó©
+ q1 (1 − q2 )EF max{x, x̂} + q2 (1 − q1 )EF min{x, x̂} + q1 q2 EF (x) .
We can rewrite the term in brackets in the same way we rewrote the agent’s payoff above
to obtain
X
F ∈F
σ(F )EF (x) =
X
¶
î
ó©
σ(F ) α̂EF (x) + (1 − α̂) (1 − q̂1 )x̂ + q̂1 EF max{x, x̂} ,
F ∈F
which is the same equation that would define x̂ given σ in the game (F, α̂, q̂1 , 0).
A similar substitution and rearrangement shows the result for q2 ≥ q1 .
This result also holds for arbitrary correlation between the event that the agent
receives evidence and the event that the challenger does. To see this, let pb be the
probability that both have evidence, p1 the probability that only the agent has evidence,
p2 the probability that only the challenger has evidence, and pn the probability that
neither has evidence. So we now reinterpret q1 to be the marginal probability that the
agent has evidence — that is, q1 = p1 + pb — and reinterpret q2 analogously. It is easy
to see that our argument that the challenger will reveal any x he observes with x ≤ x̂
and that the agent will reveal any x ≥ x̂ does not rely on any correlation assumption, so
the agent’s payoff as a function of F and x̂ is now
î
ó
αEF (x) + (1 − α) pn x̂ + p2 EF min{x, x̂} + p1 EF max{x, x̂} + pb EF (x) .
If we again substitute from equation (5), we obtain
αEF (x) + (1 − α) [(pb + p2 )EF (x) + (pn + p2 )x̂ + (p1 − p2 )EF max{x, x̂}] .
But p2 + pb = q2 , pn + p2 = 1 − pb − p1 = 1 − q1 , and p1 − p2 = q1 − q2 . Substituting these
expressions, we can rearrange to obtain equation (6) and complete the proof exactly as
above.
We can use Theorem 7 to extend Theorems 3 and 5 to this setting. To see this, note
that the former theorem tells us that the worst possible payoff for the agent in (F, α, q1 , 0)
is the first–best payoff times
α + (1 − α)q1
.
α + (1 − α)q1 (2 − q1 )
23
Reinterpret this as our “translation” of a game (F, α, q1 , q2 ) where q1 > q2 . In other
words, we can treat this lower bound as
α̂ + (1 − α̂)q̂1
α̂ + (1 − α̂)q̂1 (2 − q̂1 )
where α̂ = α + (1 − α)q2 and q̂1 = (q1 − q2 )/(1 − q2 ). We can substitute in and rearrange
to obtain a lower bound as a function of (α, q1 , q2 ) when q1 > q2 of
(1 − q2 )[α + (1 − α)q1 ]
.
α + (1 − α)q1 (2 − q1 ) − q2
Similar reasoning gives a lower bound when q2 > q1 of
α + (1 − α)q1
.
α + (1 − α)q2
These bounds reinforce the message of Theorem 7 in that both expressions equal 1
when q1 = q2 if either α > 0 or q1 > 0. Thus for any nondegenerate game, we obtain the
first–best when q1 = q2 .
It is intuitive and not hard to see that the properties of R discussed earlier for the
cases q1 = 0 and q2 = 0 hold in general. Specifically, the worst–case payoff is increasing
in α and hence is minimal at α = 0. If q2 > q1 , then it is decreasing in q2 , while if
q1 > q2 , it is non–monotonic in q1 . In addition, we now can see that if qi > qj , then R is
continuously increasing in qj up to the first best when qj = qi . That is, making the less
informed player more equally informed is beneficial. Hence the worst case is that the less
informed player has no information at all.
7
Discussion
The simplicity of the model implies that many important factors are omitted, but also
suggests that the force pointed to here is very basic. As stressed in the introduction, the
key observation is that to the extent that the agent can control the flow of information,
he has incentives to take excessive risks since he can (temporarily) hide bad outcomes. To
the extent that hostile forces control the flow of information, the agent has the opposite
incentive, namely to avoid risk to an excessive degree.
In this concluding section, we briefly discuss some of the potentially interesting factors
omitted here. First, we have omitted the possibility of “noise” in the disclosure process. It
is natural to wonder if our results are robust to the possibility that the evidence disclosed
by either the agent or challenger is a noisy signal of x rather than the realization x itself.
24
To see why one might suspect nonrobustness, consider the model where only the
agent may have evidence and suppose that there are two projects, F and G, where F
yields x = 2 with certainty and G gives x = 0 or x = 3, each with probability 1/2.
For any sufficiently small α and any q1 ∈ (0, 1), in the model without noise, it is never
an equilibrium for the agent to choose F . However, now suppose that the evidence the
agent might obtain in the disclosure stage is noisy. Specifically, suppose there is a set
of signals, say Σ, and that the distribution over signals received by the agent is a full
support distribution which depends on the true outcome. That is, if the true outcome is
x, then the distribution over signals is ψ(· | x) and this distribution has full support on
Σ for any x. Then it is always an equilibrium for the agent to choose F . If the observer
expects the agent to choose F , then he expects x to equal 2 and his belief will not change
regardless of the signal the agent shows him, if any. Hence the agent has no incentive to
deviate.
On the other hand, it is easy to see that this example relies critically on the degeneracy
of the chosen project. In fact, if we assume that all projects have the same support, then
the discontinuity at zero noise disappears. To see this, think of the observer’s belief about
the project chosen by the agent as giving the observer’s prior belief over x. For any full
support “prior,” sufficiently precise signals will generate a posterior belief close to the
true realization of the outcome. Thus if all projects have the same support, the fact
that the observer’s prior would be, in a sense, wrong when the agent deviates will not
prevent the observer from assigning probability close to the 1 to the true outcome if the
agent discloses a sufficiently precise signal. Consequently, the set of equilibria for “small
noise” and for “zero noise” will necessarily be “close.”16 While our analysis is therefore
robust with respect to small amounts of noise under this full–support assumption, the
introduction of noise may introduce new issues and effects worth exploring.
A second simplification is our assumption that all projects are equally “transparent”
in the sense that the probability that the agent or challenger receives evidence to disclose
is independent of the project chosen by the agent. It is not difficult to give analogs of
our main results for a model where the probability with which evidence is received varies
with the project. For example, it is easy to show that if two nondegenerate projects are
identical except that one has a larger probability that the agent receives evidence, then
the project with the smaller probability of receiving evidence must have zero probability
in any equilibrium, a result analogous to Theorems 1 and 2. One can also give worst
case results analogous to Theorem 3. For example, it is not difficult to show that if q
can vary across projects, then for any set of feasible projects and any α, the agent’s
16
It is worth noting that we could also add noise to the model in way which obviously has no effect on
our results. Specifically, suppose that the realized outcome is the signal drawn in the disclosure phase
(whether this is observed or not) plus an independent, mean zero, random variable. In this case, the
best estimate of the outcome conditional on the disclosure of a signal realization of x is simply x, so
none of our analysis changes at all. We thank Andy Skrzypacz for pointing this out.
25
equilibrium payoff must be at least α times the first–best payoff and that this lower
bound is approximately achievable. Both of these results are stated more formally and
proved in Appendix D.
A different way to think about variation in the transparency of projects is to suppose
that the agent can take actions which determine the probability that he or the challenger
receive evidence. There are a number of delicate modeling questions here. Are the agent’s
actions regarding transparency observable? If so, he may have the ability to commit to
a q1 . In this case, at least if these actions are costless, he would commit to q1 = 1
and achieve the first–best outcome. If his actions aren’t observable but are costless, he
still has an incentive to choose q1 = 1 since this ensures he can disclose if he wishes
to do so. On the other hand, if his actions are unobserved and costly, things are more
complex, particularly if the challenger can also choose actions which affect his probability
of receiving evidence.
Given the severe inefficiency of equilibria in this environment, it is natural to ask
whether players would find ways to improve the outcomes by some richer incentive devices. In some cases, this seems difficult or impossible — e.g., in the voting example.
There it seems that the best one can do is to give equal access to information to the
challenger and incumbent (something that presumably a free press can help maintain).
In other environments, contracting may help. For example, suppose the agent is the
manager of a firm and the observer is the stock market. Then it seems natural to expect
the firm’s stockholders to alter the agent’s compensation in order to induce more efficient
behavior. Intuitively, the model implies that inefficiency results in part from the fact that
the manager’s payoff is increasing in the “short run” stock price — i.e., the stock price
before the outcome of the project is revealed to all. If his payoff instead depended only on
the “long run” stock price — i.e., the realization of x — the outcome would be first–best.
As has been noted in the literature,17 there are good reasons for expecting managerial
compensation to depend positively on both short and long run stock prices. First, if
the long run is indeed long, the manager requires compensation in the short run too.
Given limited liability, it seems implausible that he can be forced to repay short run
compensation if the realization of the project turns out to be poor in the long run.
Second, there is an issue as to whether stockholders can commit to not rewarding short
run stock prices. To see the point, suppose that stockholders may need to sell their
holdings in the short run and hence care about the short run stock price.18 If the
manager has positive news in the second period, then they would be better off at this
point if he would disclose it. Hence even if the original contract for the manager did not
reward him for a high short run stock price, the stockholders would have an incentive
17
See, for example, the discussion in Stein (1989) or Edmans, Heinle, and Huang (2013).
This formulation is common in the literature. See, for example, Diamond and Verrecchia (1991) or
Gigler, Kanodia, Sapra, and Venugopalan (2013).
18
26
to renegotiate the contract after the project choice is made. Of course, if the manager
anticipates this, it is as if the original contract depended on the short run price. Optimal
contracting in such an environment is a natural next step to consider.
27
Appendix
A
Proof of Theorem 3
Consider any game (F, α, q1 , 0). Since the conclusion that R(0, 0, 0) = 0 was shown in
the text, we focus here only on nondegenerate games so either α > 0 or q1 > 0 (or both).
It is easy to see that R(1, q1 , 0) = 1. If α = 1, the agent’s payoff from choosing F is
EF (x), independently of the strategy of the observer. Hence he must maximize this and
so his payoff must be the first–best. For the rest of this proof, assume α < 1.
It is also not hard to show that R(α, 1, 0) = 1. To see this, suppose q1 = 1 but we
have an equilibrium in which the agent’s payoff is strictly below the first–best. Then
the agent could deviate to any first–best project and always disclose the outcome. Since
q1 = 1, this ensures the agent a payoff equal to the first–best, a contradiction. Since
equilibria always exist, we see that R(α, 1, 0) = 1. For the rest of this proof, we assume
q1 < 1.
For a fixed x̂, the agent’s payoff to choosing F is
αEF (x) + (1 − α)[(1 − q1 )x̂ + q1 EF max{x, x̂}].
As shown in the text, EF max{x, x̂} = EF (x) +
R x̂
0
(7)
F (x) dx, so we can rewrite this as
(α + (1 − α)q1 )EF (x) + (1 − α)(1 − q1 )x̂ + (1 − α)q1
Z x̂
0
F (x) dx.
Fix an equilibrium mixed strategy for the agent σ and the associated x̂. Let U =
σ(F 0 )EF 0 (x), so this is the agent’s expected payoff in the equilibrium. Let F be
any project in the support of the agent’s mixed strategy such that EF (x) ≤ U and let G
be any other feasible project. Then we must have
P
F 0 ∈F
(α + (1 − α)q1 )EG (x) + (1 − α)q1
Z x̂
0
G(x) dx
≤ (α + (1 − α)q1 )EF (x) + (1 − α)q1
Z x̂
0
F (x) dx.
Since G(x) ≥ 0, this implies
(α + (1 − α)q1 )EG (x) ≤ (α + (1 − α)q1 )EF (x) + (1 − α)q1
28
Z x̂
0
F (x) dx.
Define z = 0x̂ F (x) dx/x̂. It is not hard to use equation (2) to show that q1 < 1 implies
x̂ > 0, so this is well–defined.19 Since F (x) ∈ [0, 1], we must have z ∈ [0, 1]. Then we
can rewrite this equation as
R
(α + (1 − α)q1 )EG (x) ≤ (α + (1 − α)q1 )EF (x) + (1 − α)q1 z x̂.
(8)
Since F is in the support of the agent’s equilibrium mixed strategy, we must have
(α + (1 − α)q1 )EF (x) + (1 − α)(1 − q1 )x̂ + (1 − α)q1 z x̂ = U,
so
x̂ =
U − (α + (1 − α)q1 )EF (x)
.
(1 − α)[1 − q1 + zq1 ]
Substituting into equation (8) gives
ñ
ô
U − (α + (1 − α)q1 )EF (x)
(α + (1 − α)q1 )EG (x) ≤ (α + (1 − α)q1 )EF (x) + q1 z
. (9)
1 − q1 + zq1
Recall that U ≥ EF (x), so U ≥ (α + (1 − α)q1 )EF (x). Hence q1 ≥ 0 implies that the
right–hand side is weakly increasing in z. Hence
(α + (1 − α)q1 )EG (x) ≤ (α + (1 − α)q1 )EF (x) + q1 [U − (α + (1 − α)q1 )EF (x)]
or
[α + (1 − α)q1 ]EG (x) ≤ U q1 + EF (x)(α + (1 − α)q1 )(1 − q1 ).
Since the term multiplying EF (x) is positive, the fact that EF (x) ≤ U implies
(α + (1 − α)q1 )EG (x) ≤ U [q1 + (α + (1 − α)q1 )(1 − q1 )] .
Hence, taking G to be a first–best project,
ñ
U ≥U
FB
ô
α + (1 − α)q1
.
α + (1 − α)q1 (2 − q1 )
To show that this bound is tight, consider the following example. Suppose F =
{F, G}. Assume F is a a distribution putting probability 1 − p on 0 and p on U/p, so
EF (x) = U , for some p ∈ (0, 1) and U > 0. Let G be a distribution putting probability
19
To see this, suppose x̂ = 0. Then equation (2) implies that either q1 = 1 or EF (x) = 0 for all F in
the support of the agent’s mixed strategy. Since q1 < 1 by assumption, this implies U = 0. But this is
not possible. The agent can deviate to any project with a strictly positive mean (since there are at least
two projects, such a project must exist) and always show the outcome. Since either α > 0 or q1 > 0 or
both, the agent would gain by such a deviation.
29
1 on x∗ for some x∗ > U . Note that EF (x) = U < x∗ = EG (x), so U F B = x∗ . We
will characterize a situation where F is a pure strategy equilibrium and show that this
establishes the bound. Note that ifR F is chosen with probability
1 in equilibrium, then
R x̂
x̂
∗
we must have x̂ < U < x . Hence 0 G(x) dx = 0 and 0 F (x) dx = (1 − p)x̂. Hence F
is optimal for the agent iff equation (8) holds at EG (x) = x∗ , EF (x) = U , and z = 1 − p.
We can also solve for x̂ exactly as above with z = 1 − p and EF (x) = U . Therefore, from
equation (9), this is an equilibrium iff
ñ
Ç
1 − (α + (1 − α)q1 )
(α + (1 − α)q1 )x ≤ U α + (1 − α)q1 + q1 (1 − p)
1 − q1 + (1 − p)q1
åô
∗
.
Tedious algebra leads to
Ç
U ≥x
∗
å
(α + (1 − α)q1 )(1 − q1 + (1 − p)q1 )
.
(α + (1 − α)q1 )(1 − q1 ) + (1 − p)q1
Fix p and choose x∗ so that this holds with equality. (It is immediate that the resulting
x∗ is necessarily larger than U , as assumed.) For p arbitrarily close to 0, we obtain an
example where
Ç
U ≈U
FB
α + (1 − α)q1
(α + (1 − α)q1 )(1 − q1 ) + q1
Hence
R(α, q1 , 0) =
å
Ç
=U
FB
å
α + (1 − α)q1
.
α + (1 − α)q1 (2 − q1 )
α + (1 − α)q1
.
α + (1 − α)q1 (2 − q1 )
It is not hard to show that 1/R is concave in q1 and that the first–order condition for
maximization of 1/R holds uniquely at
√
α
√ .
q1 =
1+ α
Thus R is uniquely minimized at this q1 . Substituting this value of q1 into R and rearranging yields
√
1+ α
min R(α, q1 , 0) =
,
q1 ∈[0,1]
2
as asserted.
B
Proof of Theorem 5
Again, nondegeneracy implies that either α > 0 or q2 > 0 or both. Just as in the proof
of Theorem 3, the result that we obtain the first–best when α = 1 is straightforward, so
30
we assume throughout this proof that α < 1. The case of α = 0 is also straightforward.
To see this, suppose there is a distribution F ∈ F which is degenerate at 0. Suppose the
observer believes the agent chooses this distribution and the challenger never shows any
strictly positive x. Then since α = 0, no deviation by the agent can achieve a strictly
positive payoff. No matter what the agent does, the observer’s belief is that x = 0, so
the agent’s payoff is zero. Hence this is an equilibrium, establishing that R(0, 0, q2 ) = 0
for any q2 . Hence for the rest of this proof, we assume α ∈ (0, 1).
Given that q1 = 0, we can write the agent’s payoff given x̂ and a choice of project F
as
αEF (x) + (1 − α)(1 − q2 )x̂ + (1 − α)q2 EF min{x, x̂}.
Since EF min{x, x̂} =
R x̂
0
[1 − F (x)] dx, we can rewrite this as
αEF (x) + (1 − α)(1 − q2 )x̂ + (1 − α)q2
Z x̂
0
[1 − F (x)] dx.
So fix an equilibrium mixed strategy for the agent σ and the associated x̂. Again, let
P
U be the agent’s expected payoff — that is, U = F 0 ∈F σ(F 0 )EF 0 (x). Let F be a project
in the support of the agent’s mixed strategy satisfying EF (x) ≤ U and let G be any other
feasible project. Then we must have
αEG (x) + (1 − α)q2
Z x̂
0
[1 − G(x)] dx ≤ αEF (x) + (1 − α)q2
Z x̂
0
[1 − F (x)] dx.
Since G(x) ≤ 1, this implies
αEG (x) ≤ αEF (x) + (1 − α)q2
Z x̂
0
[1 − F (x)] dx.
Define z = 0x̂ [1 − F (x)] dx/x̂. One can use equation (2) and α > 0 to show that x̂ > 0
so this is well–defined.20 As in the proof of Theorem 3, F (x) ∈ [0, 1] implies z ∈ [0, 1].
Then we can rewrite this equation as
R
αEG (x) ≤ αEF (x) + (1 − α)q2 z x̂.
(10)
Because F is in the support of the agent’s equilibrium mixed strategy, we must have
αEF (x) + (1 − α)(1 − q2 )x̂ + (1 − α)q2 z x̂ = U,
20
To see this, suppose x̂ = 0. From equation (2), this implies that
X
X
σ(F 0 )EF 0 (x) = q2
σ(F 0 )EF 0 min{x, 0} = 0.
F 0 ∈F
F 0 ∈F
Hence the agent’s mixed strategy must put probability 1 on a degenerate distribution at 0 and so U = 0.
Since α > 0, the agent can deviate to any other project (which must have a strictly positive mean) and
be strictly better off even if the challenger never discloses anything.
31
so
x̂ =
U − αEF (x)
.
(1 − α)(1 − q2 + zq2 )
Substituting into equation (10) gives
ñ
ô
U − αEF (x)
αEG (x) ≤ αEF (x) + q2 z
.
1 − q2 + zq2
(11)
By assumption, U ≥ EF (x), so U ≥ αEF (x). Hence the right–hand side is weakly
increasing in z, so this implies
αEG (x) ≤ αEF (x) + q2 [U − αEF (x)]
or
αEG (x) ≤ q2 U + α(1 − q2 )EF (x) ≤ U [q2 + α(1 − q2 )] = U [α + (1 − α)q2 ].
Hence, taking G to be a first–best project,
ñ
U ≥U
FB
ô
α
.
α + (1 − α)q2
To see that the bound is tight, let F be a degenerate distribution at x∗ and suppose
we have an equilibrium where the agent chooses F . Clearly, then, x̂ = U = x∗ . Let G
put probability 1 − p on 0 and p on y/p where y > x∗ for some p ∈ (0, 1). Note that
EG (x) = y. Assume F and G are the only feasible projects. Then this is an equilibrium
if
αy + (1 − α)q2 [(1 − p)(0) + px̂] ≤ (α + (1 − α)q2 )x̂.
Since x̂ = U , we can rewrite this as
αy ≤ U [α + (1 − α)(1 − p)q2 ] .
Fix any p ∈ (0, 1) and choose y so that this holds with equality. Since the resulting y
satisfies y ≥ U , we have U F B = y. So this gives an example where
ñ
U =U
FB
ô
α
.
α + (1 − α)(1 − p)q2
As p ↓ 0, the right–hand side converges to α/[α + (1 − α)q2 ]. Hence we can get arbitrarily
close to the stated bound, so
R(α, 0, q2 ) =
α
.
α + (1 − α)q2
The last two statements of the theorem follow directly.
32
C
Comparative Statics Example
Suppose there are three feasible projects, F1 , F2 , and F3 . Project Fi gives a “high
outcome” hi with probability pi and a “low outcome” `i otherwise. The specific values
of hi , `i , and pi are given in the table below.
F1
F2
F3
hi
`i
pi
µi
964
532 1/2 748
5904/7 0 7/8 738
1737/2 171 4/5 729
In the table, µi = EFi (x). Note that F1 is the first–best project, F2 is second best,
and F3 worst. Simple calculations show the range of α’s for which it is a pure strategy
equilibrium for the agent to choose Fi for each i. For each of the three projects, there
is a nonempty range of α’s where it is chosen in equilibrium. Similarly, for each pair of
projects, there is a nonempty range of α’s where that pair is the support of the agent’s
mixed strategy.
In the case where the agent randomizes between projects F1 and F2 or between F1 and
F3 , the agent’s equilibrium payoff decreases with α. On the other hand, the equilibrium
payoff when the agent randomizes between F2 and F3 is increasing in α.
To see the intuition, consider the case where the agent randomizes between F1 and
F2 . As α increases, if x̂ is fixed, the agent would switch to F1 since he now cares more
about the outcome of the project and F1 has the higher expected outcome. So x̂ must
adjust to deter this deviation. Which way do we need to adjust x̂ to make the agent
indifferent again? Note that F2 has a much higher chance of having a good outcome to
show than F1 . Thus if x̂ declines, this pushes the agent toward F2 . Hence the adjustment
that restores indifference is reducing x̂. To reduce x̂, we must make the observer more
pessimistic about the outcome. This means we must reduce the probability that the
agent picks F1 , lowering the agent’s equilibrium payoff. Similarly, note that F3 gives its
high outcome with higher probability than F1 , so similar intuition applies here. On the
other hand, in comparing F2 and F3 , it is F2 , the better of the two projects, which has
the higher chance of the high outcome. Hence the opposite holds in this case.
The figure below shows the equilibrium payoffs as a function of α. Note that, as
asserted, the equilibrium payoffs for two of the three mixed strategy equilibria are decreasing in α. Note also that the payoff to the worst equilibrium is decreasing in α for
α between 1/4 and 1/3. Finally, note that if we focus only on pure strategy equilibria, the worst equilibrium payoff is decreasing in α as we move from the range where
α ∈ [5/24, 1/3] to α > 1/3.
33
F1
748
F1 + F3
741
F1 + F2
F2
738
F2 + F3
729
F3
1 5
5 24
1
4
4
15
1
3
Figure 3: Equilibrium Payoffs
34
α
D
q Varying Across Projects
In this section, we consider a variation on our model where the challenger never has
evidence and the probability the agent has evidence depends on the project he selects.
Here we denote a project by the pair (F, qF ) where F is a probability distribution over
outcomes x and qF is the probability the agent receives evidence he can disclose. We
show two results for this model which are analogs for Theorems 2 and 3.
Theorem 8. Suppose there are feasible projects (F, qF ) and (G, qG ) where F = G, F is
nondegenerate, and qG > qF . Then if α < 1, project (F, qF ) is chosen with zero probability
in any equilibrium.
Proof. Suppose to the contrary that there is an equilibrium in which (F, qF ) is chosen
with strictly positive probability. Then we must have
αEF (x) + (1 − α)(1 − qF )x̂ + (1 − α)qF EF max{x, x̂}
≥ αEG (x) + (1 − α)(1 − qG )x̂ + (1 − α)qG EG max{x, x̂}
where x̂ is the observer’s belief if the agent does not disclose any evidence. Since F = G,
this implies
qF [EF max{x, x̂} − x̂] ≥ qG [EF max{x, x̂} − x̂].
Since qG > qF , this requires x̂ ≥ x̄F where x̄F is the upper bound of the support of F .
Given this, the payoff to F in equilibrium is αEF (x) + (1 − α)x̂ < x̂. The inequality
follows from x̂ ≥ x̄F and is strict because F is nondegenerate by assumption. But it is
easy to show that the agent’s equilibrium payoff is
X
σ(F 0 )EF 0 (x) ≥ x̂,
F0
where σ is the agent’s mixed strategy. Hence the agent’s equilibrium payoff strictly
exceeds the payoff to project (F, qF ), a contradiction.
Theorem 9. For any set of feasible projects, any α ∈ [0, 1], and any equilibrium, the
agent’s payoff is at least α times the first–best payoff. Furthermore, there exists a set
of feasible projects and an equilibrium such that the agent’s payoff equals α times the
first–best.
Proof. To show the bound, fix any set of feasible projects, any α, and any equilibrium.
Let U be the agent’s payoff in the equilibrium and let x̂ be the belief in response to no
disclosure in the equilibrium. Let (F, qF ) be any first–best project. Then
U ≥ αEF (x) + (1 − α)qF EF max{x, x̂} + (1 − α)(1 − qF )x̂
≥ αEF (x)
35
where the second inequality uses the fact that x ≥ 0 with probability 1. Hence U is at
least α times the first–best payoff.
To see that this is attainable, fix any y ≥ 0 and any U ∈ [αy, y). Let the feasible set
of projects consist of two projects, (F, 0) and (G, 1) where F yields y with probability 1
and G yields 2U with probability 1/2 and 0 otherwise. Clearly, (F, 0) is the first–best
project. However, it is easy to see that it is an equilibrium for the agent to choose project
(G, 1). To see this, suppose it is the project the observer expects. Then x̂ must satisfy
1
1
U = x̂ + (2U ),
2
2
so x̂ = 0. Hence if the agent were to deviate to project (F, 0), his payoff would be
αy + (1 − α)(0). Since U ≥ αy, the agent has no incentive to deviate from (G, 1), so
this is an equilibrium. In particular, this construction gives an equilibrium even when
U = αy, showing there is an equilibrium with payoff equal to α times the first–best.
36
References
[1] Acharya, V., P. DeMarzo, and I. Kremer, “Endogenous Information Flows and the
Clustering of Announcements,” American Economic Review, 101, 2011, 2955–2979.
[2] Beyer, A., and I. Guttman, “Voluntary Disclosure, Manipulation and Real Effects,”
Stanford University working paper, 2012.
[3] Bond, P., and I. Goldstein, “Government Intervention and Information Aggregation
by Prices,” University of Pennsylvania working paper, 2014.
[4] Diamond, D., and R. Verrecchia, “Disclosure, Liquidity, and the Cost of Capital,”
Journal of Finance, 46, September 1991, 1325–1359.
[5] Dye, R. A., “Disclosure of Nonproprietary Information,” Journal of Accounting Research, 23, 1985a, 123–145.
[6] Dye, R. A., “Strategic Accounting Choice and the Effect of Alternative Financial
Reporting Requirements,” Journal of Accounting Research, 23, 1985b, 544–574.
[7] Edmans, A., M. Heinle, and C. Huang, “The Real Costs of Disclosure,” Wharton
working paper, May 2013.
[8] Fishman, M. and K. Hagerty, “The Optimal Amount of Discretion to Allow in
Disclosure,” Quarterly Journal of Economics, 105, May 1990, 427–444.
[9] Forges, F. and F. Koessler, “Communication Equilibria with Partially Verifiable
Types,” Journal of Mathematical Economics, 41, November 2005, 793–811.
[10] Forges, F. and F. Koessler, “Long Persuasion Games,” Journal of Economic Theory,
143, November 2008, 1–35.
[11] Gao, P., and P. Liang, “Informational Feedback Effect, Adverse Selection, and the
Optimal Disclosure Policy,” Journal of Accounting Research, 51, December 2013,
1133–1158.
[12] Gigler, F., C. Kanodia, H. Sapra, and R. Venugopalan, “How Frequent Financial
Reporting Can Cause Managerial Short–Termism: An Analysis of the Costs and
Benefits of Increasing Reporting Frequency,” University of Minnesota working paper,
September 2013.
[13] Glazer, J., and A. Rubinstein, “On Optimal Rules of Persuasion,” Econometrica,
72, November 2004, 1715–1736.
[14] Glazer, J., and A. Rubinstein, “A Study in the Pragmatics of Persuasion: A Game
Theoretical Approach,” Theoretical Economics, 1, December 2006, 395–410.
37
[15] Grossman, S. J., “The Informational Role of Warranties and Private Disclosures
about Product Quality,” Journal of Law and Economics, 24, 1981, 461–483.
[16] Guttman, I., I. Kremer, and A. Skrzypacz, “Not Only What but also When: A Theory of Dynamic Voluntary Disclosure,” American Economic Review, forthcoming,
2013.
[17] Jensen, M., and W. Meckling, “Theory of the Firm: Managerial Behavior, Agency
Costs and Ownership Structure,” Journal of Financial Economics, 3, October 1976,
305–360.
[18] Jung, W., and Y. Kwon, “Disclosure When the Market is Unsure of Information
Endowment of Managers,” Journal of Accounting Research, 26, 1988, 146–153.
[19] Kanodia, C., and A. Mukherji, “Real Effects of Separating Investment and Operating
Cash Flows,” Review of Accounting Studies, 1, 1996, 51–71.
[20] Kanodia, C., H. Sapra, and R. Venugopalan, “Should Intangibles Be Measured:
What Are the Economic Trade–Offs?” Journal of Accounting Research, 42, March
2004, 89–120.
[21] Koutsoupias, E., and C. Papadimitriou, “Worst–case Equilibria,” in Proceedings
of the 16th Annual Symposium on Theoretical Aspects of Computer Science, 1999,
pp. 404–413.
[22] Kreps, D., and R. Wilson, “Sequential Equilibrium,” Econometrica, 50, July 1982,
863–894.
[23] Lipman, B., and D. Seppi, “Robust Inference in Communication Games with Partial
Provability,” Journal of Economic Theory, 66, August 1995, 370–405.
[24] Milgrom, P., “Good News and Bad News: Representation Theorems and Applications,” Bell Journal of Economics, 12, 1981, 350–391.
[25] Mas-Colell, A., M. Whinston, and J. Green, Microeconomic Theory, New York:
Oxford University Press, 1995.
[26] Okuno-Fujiwara, M., A. Postlewaite, and K. Suzumura, “Strategic Information Revelation,” Review of Economic Studies, 57, January 1990, 25–47.
[27] Roughgarden, T., Selfish Routing and the Price of Anarchy, MIT Press, 2005.
[28] Shin, H. S., “The Burden of Proof in a Game of Persuasion,” Journal of Economic
Theory, 64, October 1994, 253–264.
[29] Shin, H. S., “Disclosures and Asset Returns,” Econometrica, 71, January 2003,
105–133.
38
[30] Stein, J., “Efficient Capital Markets, Inefficient Firms: A Model of Myopic Corporate
Behavior,” Quarterly Journal of Economics, 101, November 1989, 655–669.
[31] Verrecchia, R. E., “Discretionary Disclosure,” Journal of Accounting and Economics,
5, 1983, 179–194.
39